Transformers have been widely adopted in the field of hyperspectral image (HSI) classification. However, a significant drawback of transformers lies in their excessive number of parameters and the high computational overhead. To address this challenge, we propose a Channel-Reduced Transformer (CRFormer) for HSI classification. In an effort to enhance computational efficiency, we first introduce a cross-region tokenization (CRT) approach. This method effectively shortens the sequence length input to the transformer, thereby alleviating the computational burden. Additionally, we propose a channel-reduced multi-head self-attention (CR-MHSA) module. This module operates on only half of the input channels while still attaining comparable or even superior results. Experimental results conducted on three benchmark datasets demonstrate that our proposed method not only achieves superior classification accuracy but also significantly reduces computational complexity compared to other transformer-based approaches.

Deep learning technology has rapidly advanced in the interpretation of polarimetric synthetic aperture radar (PolSAR) images in recent years. However, deep learning methods in PolSAR image interpretation primarily rely on a significant volume of labeled data to make precise predictions, while disregarding the potential physical features of PolSAR. To solve the problem, a deep fusion network is proposed in this letter, which can effectively utilize the complementary characteristics between amplitude and physical features of PolSAR images to enhance the interpretability of the network and improve PolSAR image classification performance. In addition, an improved feature pyramid network (IFPN) and a learnable feature fusion module (LFFM) were proposed to autonomously learn the required fused feature information and avoid the process of feature selection. Finally, the spectral features of PolSAR data are fused to further enhance the discriminability of the features extracted by the proposed network and improve the classification accuracy of the proposed method. In addition, to verify the effectiveness of the proposed method, two real PolSAR datasets were used. The experimental results demonstrate that the proposed method achieves higher accuracy, even with a limited number of labeled samples.

The global modeling capability of vision transformer (ViT) has been well proven in the field of hyperspectral image (HSI) classification. However, ViT does not have the excellent local feature extraction capability compared with the convolutional neural network (CNN). Therefore, early-stage convolutions are often used to enhance ViT’s local representation ability. However, directly applying convolutions on high-dimensional HSI data increases computational overhead. Moreover, recent researches have observed that ViT may suffer from high redundancy in capturing multihead self-attention (MHSA). To address the above issues, we propose a multiscale super token transformer (MSSTT) model for HSI classification. We use a divide-and-conquer strategy to extract local features and global dependencies of HSI data at multiple granularities. Specifically, our proposed model incorporates two branches: a multiscale convolution (MSConv) branch that uses various convolutional kernels to extract diverse local features and a multiscale super token attention (MSSTA) branch for capturing global features with low redundancy. Finally, comparative experimental results with advanced methods show that the proposed MSSTT possesses better classification performance. On the Salinas (SA), Pavia University (PU), and Kennedy Space Center (KSC) datasets, the overall accuracies (OAs) of our MSSTT are 98.47%, 98.47%, and 99.38%, respectively. Code will be released athttps://github.com/zhangtaizheng/MSSTT.

Contrastive learning with learnable examples performs outstandingly in data representation. However, when dealing with hard samples, instance-level alignment with excessive uniformity may descend into trivial clusters, especially when confronted with inter-class similarity and intra-class diversity in hyperspectral images. To solve this problem, we regard prototypical contrastive learning as tracing the potential probability density distribution. Then, a novel pre-training method, Learnable Sparse Contrastive Sampling (LSCoSa), is proposed for discriminative representation learning, containing sparse positive sampling and multiple positives learning. Specifically, on the basis of cooperative-adversarial contrastive learning, we first exert a KL divergence regularizer on the average activation probability of the prototypes, suppressing fake density prototypes for sparse positive sampling. Furthermore, we propose multiple positives learning, in which the top-k potential positives are retrieved and dynamically weighted for contrastive supervision, to avoid trivial clusters and cover satisfying semantic variations. Comprehensive experiments on three HSI benchmark datasets demonstrate that LSCoSa achieves significant advantages over other HSIC methods. The code is available at https://github.com/sakurashine/LSCoSa.

Recently, convolutional neural networks (CNNs) show excellent performance on the hyperspectral image (HSI) classification tasks. However, traditional CNNs usually have insufficient feature discrimination and a large number of network parameters. In response to the above problems, a residual dense asymmetric convolutional network (RDACN) for HSI classification is proposed in this paper. Firstly, we de-sign a novel residual dense asymmetric convolutional block to effectively leverage the information of the previous layers. Moreover, the block adopts two feature fusion methods of addition and channel stacking to capture discriminative hyperspectral feature. Secondly, the ordinary square convolutional kernel is replaced with the asymmetric convolutional kernels, which can reduce CNN parameters. Finally, experimental results on three well-known hyperspectral datasets show that RDACN achieves competitive classification performance compared with the state-of-the-art CNNs.

Convolutional neural networks (CNNs) showed impressive performance for hyperspectral image (HSI) classification. Nevertheless, convolutional layers contain massive parameters, which restrict the deployment of CNNs on satellite and airborne platforms with limited storage and computing resources. In this letter, we propose a lightweight spectral-spatial convolution module (LS2CM) as an alternative to the convolutional layer. The proposed LS2CM can greatly reduce network parameters and computational complexity in terms of multiply-accumulate operations (MACs) while maintaining or even improving the classification performance. Furthermore, it is a plug-and-play component and can be used to upgrade existing CNN-based models for HSI classification. Experimental results on two benchmark HSI data sets demonstrate that the proposed LS2CM achieves competitive results in comparison with other state-of-the-art methods.

Hyperspectral image (HSI) can provide rich spectral information which can be helpful for accurate classification in many applications. Yet, incorporating spatial information in the classification process can improve the classification accuracy even further. Existing convolutional neural network (CNN) usually only focuses on local features in hyperspectral cubes, whereas the burgeoning vision transformer (ViT) is interested in global features in HSIs. In this letter, we propose a deep aggregated framework for HSI classification called convolution transformer mixer (CTMixer) to combine the advantages of the above two paradigms effectively. A group parallel residual block is firstly applied to capture local spectral-spatial features in the HSI patches. Secondly, a double-branch structure, consisting of the CNN and transformer branches, is developed to capture local-global hyperspectral features. Finally, to achieve an elegant combination of CNN and ViT, a novel local-global multi-head self-attention mechanism is proposed by introducing convolution operations in the multi-head self-attention mechanism to further improve the classification accuracy. Extensive experiments demonstrate that the CTMixer achieves competitive classification results on several common HSI datasets compared with other state-of-the-art networks. The source code for this work will be available at https://github.com/ZJier/CTMixer.